ci_inline int
ci_udp_recvmsg_socklocked_spin(ci_udp_iomsg_args* a,
ci_netif* ni, ci_udp_state* us,
struct recvmsg_spinstate* spin_state)
{
ci_uint64 now_frc;
int intf_i;
ci_frc64(&now_frc);
if( now_frc - spin_state->start_frc < spin_state->max_spin ) {
#if CI_CFG_SPIN_STATS
ni->state->stats.spin_udp_recv++;
#endif
if( ci_netif_may_poll(ni) ) {
OO_STACK_FOR_EACH_INTF_I(ni, intf_i)
if( ci_netif_intf_has_event(ni, intf_i) && ci_netif_trylock(ni) ) {
ci_netif_poll_intf_fast(ni, intf_i, now_frc);
ci_netif_unlock(ni);
if( ci_udp_recv_q_readable(ni, us) )
return 0;
}
if( ni->state->poll_work_outstanding ||
ci_netif_need_timer_prime(ni, now_frc) )
if( ci_netif_trylock(ni) ) {
ci_netif_poll(ni);
ci_netif_unlock(ni);
}
if( ! ni->state->is_spinner )
ni->state->is_spinner = 1;
}
return OO_SPINLOOP_PAUSE_CHECK_SIGNALS(ni, now_frc,
&spin_state->schedule_frc,
us->s.so.rcvtimeo_msec,
&us->s.b, spin_state->si);
}
int ci_tcp_shutdown(citp_socket* ep, int how, ci_fd_t fd)
{
ci_sock_cmn* s = ep->s;
int rc;
if( s->b.state == CI_TCP_LISTEN )
return ci_tcp_shutdown_listen(ep, how, fd);
if( SOCK_TO_TCP(s)->snd_delegated ) {
/* We do not know which seq number to use. Call
* onload_delegated_send_cancel(). */
CI_SET_ERROR(rc, EBUSY);
return rc;
}
if( ! ci_netif_trylock(ep->netif) ) {
/* Can't get lock, so try to defer shutdown to the lock holder. */
unsigned flags = 0;
switch( s->b.state ) {
case CI_TCP_CLOSED:
case CI_TCP_TIME_WAIT:
CI_SET_ERROR(rc, ENOTCONN);
return rc;
}
if( how == SHUT_RD || how == SHUT_RDWR )
flags |= CI_SOCK_AFLAG_NEED_SHUT_RD;
if( how == SHUT_WR || how == SHUT_RDWR )
flags |= CI_SOCK_AFLAG_NEED_SHUT_WR;
ci_atomic32_or(&s->s_aflags, flags);
if( ! ci_netif_lock_or_defer_work(ep->netif, &s->b) )
return 0;
ci_atomic32_and(&s->s_aflags, ~flags);
}
if( 0 ) {
/* Poll to get up-to-date. This is slightly spurious but done to ensure
* ordered response to all packets preceding this FIN (e.g. ANVL tcp_core
* 9.18)
*
* DJR: I've disabled this because it can hurt performance for
* high-connection-rate apps. May consider adding back (as option?) if
* needed.
*/
ci_netif_poll(ep->netif);
}
rc = __ci_tcp_shutdown(ep->netif, SOCK_TO_TCP(s), how);
if( rc < 0 )
CI_SET_ERROR(rc, -rc);
ci_netif_unlock(ep->netif);
return rc;
}
static void oo_epoll2_wait(struct oo_epoll_private *priv,
struct oo_epoll2_action_arg *op)
{
/* This function uses oo_timesync_cpu_khz but we do not want to
* block here for it to stabilize. So we already blocked in
* oo_epoll_fop_open().
*/
ci_uint64 start_frc = 0, now_frc = 0; /* =0 to make gcc happy */
tcp_helper_resource_t* thr;
ci_netif* ni;
unsigned i;
ci_int32 timeout = op->timeout;
/* Get the start of time. */
if( timeout > 0 || ( timeout < 0 && op->spin_cycles ) )
ci_frc64(&start_frc);
/* Declare that we are spinning - even if we are just polling */
OO_EPOLL_FOR_EACH_STACK(priv, i, thr, ni)
ci_atomic32_inc(&ni->state->n_spinners);
/* Poll each stack for events */
op->rc = -ENOEXEC; /* impossible value */
OO_EPOLL_FOR_EACH_STACK(priv, i, thr, ni) {
if( ci_netif_may_poll(ni) && ci_netif_has_event(ni) &&
ci_netif_trylock(ni) ) {
int did_wake;
ni->state->poll_did_wake = 0;
ci_netif_poll(ni);
did_wake = ni->state->poll_did_wake;
ci_netif_unlock(ni);
/* Possibly, we've got necessary event. If true, exit */
if( did_wake ) {
op->rc = efab_linux_sys_epoll_wait(op->kepfd,
CI_USER_PTR_GET(op->events),
op->maxevents, 0);
if( op->rc != 0 )
goto do_exit;
}
}
}
/* Do we have anything to do? */
if( op->rc == -ENOEXEC ) {
/* never called sys_epoll_wait() - do it! */
op->rc = efab_linux_sys_epoll_wait(op->kepfd, CI_USER_PTR_GET(op->events),
op->maxevents, 0);
}
if( op->rc != 0 || timeout == 0 )
goto do_exit;
/* Fixme: eventually, remove NO_USERLAND stacks from this list.
* Here is a good moment: we are going to spin or block, so there are
* a lot of time. But avoid locking! */
/* Spin for a while. */
if( op->spin_cycles ) {
ci_uint64 schedule_frc;
ci_uint64 max_spin = op->spin_cycles;
int spin_limited_by_timeout = 0;
ci_assert(start_frc);
if( timeout > 0) {
ci_uint64 max_timeout_spin = (ci_uint64)timeout * oo_timesync_cpu_khz;
if( max_timeout_spin <= max_spin ) {
max_spin = max_timeout_spin;
spin_limited_by_timeout = 1;
}
}
/* spin */
now_frc = schedule_frc = start_frc;
do {
if(unlikely( signal_pending(current) )) {
op->rc = -EINTR; /* epoll_wait returns EINTR, not ERESTARTSYS! */
goto do_exit;
}
OO_EPOLL_FOR_EACH_STACK(priv, i, thr, ni) {
#if CI_CFG_SPIN_STATS
ni->state->stats.spin_epoll_kernel++;
#endif
if( ci_netif_may_poll(ni) &&
ci_netif_need_poll_spinning(ni, now_frc) &&
ci_netif_trylock(ni) ) {
ci_netif_poll(ni);
ci_netif_unlock(ni);
}
}
op->rc = efab_linux_sys_epoll_wait(op->kepfd, CI_USER_PTR_GET(op->events),
op->maxevents, 0);
if( op->rc != 0 )
goto do_exit;
ci_frc64(&now_frc);
if(unlikely( now_frc - schedule_frc > oo_timesync_cpu_khz )) {
schedule(); /* schedule() every 1ms */
schedule_frc = now_frc;
}
else
ci_spinloop_pause();
} while( now_frc - start_frc < max_spin );
/* Move priv file to the alien_ni stack.
* Should be called with the locked priv stack and socket;
* the function returns with this stack being unlocked.
* If rc=0, it returns with alien_ni stack locked;
* otherwise, both stacks are unlocked.
* Socket is always unlocked on return. */
int efab_file_move_to_alien_stack(ci_private_t *priv, ci_netif *alien_ni)
{
tcp_helper_resource_t *old_thr = priv->thr;
tcp_helper_resource_t *new_thr = netif2tcp_helper_resource(alien_ni);
ci_sock_cmn *old_s = SP_TO_SOCK(&old_thr->netif, priv->sock_id);
ci_sock_cmn *new_s;
ci_sock_cmn *mid_s;
tcp_helper_endpoint_t *old_ep, *new_ep;
int rc, i;
int pollwait_register = 0;
#if CI_CFG_FD_CACHING
oo_p sp;
#endif
OO_DEBUG_TCPH(ci_log("%s: move %d:%d to %d", __func__,
old_thr->id, priv->sock_id, new_thr->id));
/* Poll the old stack - deliver all data to our socket */
ci_netif_poll(&old_thr->netif);
/* Endpoints in epoll list should not be moved, because waitq is already
* in the epoll internal structures (bug 41152). */
if( !list_empty(&priv->_filp->f_ep_links) ) {
rc = -EBUSY;
goto fail1;
}
if( !efab_file_move_supported(&old_thr->netif, old_s) ) {
rc = -EINVAL;
goto fail1;
}
/* Lock the second stack */
i = 0;
while( ! ci_netif_trylock(alien_ni) ) {
ci_netif_unlock(&old_thr->netif);
if( i++ >= 1000 ) {
rc = -EBUSY;
goto fail1_ni_unlocked;
}
rc = ci_netif_lock(&old_thr->netif);
if( rc != 0 )
goto fail1_ni_unlocked;
}
/* Allocate a new socket in the alien_ni stack */
rc = -ENOMEM;
if( old_s->b.state == CI_TCP_STATE_UDP ) {
ci_udp_state *new_us = ci_udp_get_state_buf(alien_ni);
if( new_us == NULL )
goto fail2;
new_s = &new_us->s;
}
else {
ci_tcp_state *new_ts = ci_tcp_get_state_buf(alien_ni);
if( new_ts == NULL )
goto fail2;
new_s = &new_ts->s;
}
/* Allocate an intermediate "socket" outside of everything */
mid_s = ci_alloc(CI_MAX(sizeof(ci_tcp_state), sizeof(ci_udp_state)));
if( mid_s == NULL )
goto fail3;
OO_DEBUG_TCPH(ci_log("%s: move %d:%d to %d:%d", __func__,
old_thr->id, priv->sock_id,
new_thr->id, new_s->b.bufid));
/* Copy TCP/UDP state */
memcpy(mid_s, old_s, CI_MAX(sizeof(ci_tcp_state), sizeof(ci_udp_state)));
/* do not copy old_s->b.bufid
* and other fields in stack adress space */
mid_s->b.sb_aflags |= CI_SB_AFLAG_ORPHAN;
mid_s->b.bufid = new_s->b.bufid;
mid_s->b.post_poll_link = new_s->b.post_poll_link;
mid_s->b.ready_link = new_s->b.ready_link;
mid_s->reap_link = new_s->reap_link;
if( old_s->b.state & CI_TCP_STATE_TCP ) {
ci_tcp_state *new_ts = SOCK_TO_TCP(new_s);
ci_tcp_state *mid_ts = SOCK_TO_TCP(mid_s);
mid_ts->timeout_q_link = new_ts->timeout_q_link;
mid_ts->tx_ready_link = new_ts->tx_ready_link;
mid_ts->rto_tid = new_ts->rto_tid;
mid_ts->delack_tid = new_ts->delack_tid;
mid_ts->zwin_tid = new_ts->zwin_tid;
mid_ts->kalive_tid = new_ts->kalive_tid;
mid_ts->cork_tid = new_ts->cork_tid;
ci_ip_queue_init(&mid_ts->recv1);
ci_ip_queue_init(&mid_ts->recv2);
ci_ip_queue_init(&mid_ts->send);
ci_ip_queue_init(&mid_ts->retrans);
mid_ts->send_prequeue = OO_PP_ID_NULL;
new_ts->retrans_ptr = OO_PP_NULL;
mid_ts->tmpl_head = OO_PP_NULL;
oo_atomic_set(&mid_ts->send_prequeue_in, 0);
*new_ts = *mid_ts;
ci_pmtu_state_init(alien_ni, &new_ts->s, &new_ts->pmtus,
CI_IP_TIMER_PMTU_DISCOVER);
#if CI_CFG_FD_CACHING
sp = TS_OFF(alien_ni, new_ts);
OO_P_ADD(sp, CI_MEMBER_OFFSET(ci_tcp_state, epcache_link));
ci_ni_dllist_link_init(alien_ni, &new_ts->epcache_link, sp, "epch");
ci_ni_dllist_self_link(alien_ni, &new_ts->epcache_link);
sp = TS_OFF(alien_ni, new_ts);
OO_P_ADD(sp, CI_MEMBER_OFFSET(ci_tcp_state, epcache_fd_link));
ci_ni_dllist_link_init(alien_ni, &new_ts->epcache_fd_link, sp, "ecfd");
ci_ni_dllist_self_link(alien_ni, &new_ts->epcache_fd_link);
#endif
/* free temporary mid_ts storage */
CI_FREE_OBJ(mid_ts);
}
else {
ci_udp_state *mid_us = SOCK_TO_UDP(mid_s);
*SOCK_TO_UDP(new_s) = *mid_us;
CI_FREE_OBJ(mid_us);
}
/* Move the filter */
old_ep = ci_trs_ep_get(old_thr, priv->sock_id);
new_ep = ci_trs_ep_get(new_thr, new_s->b.bufid);
rc = tcp_helper_endpoint_move_filters_pre(old_ep, new_ep);
if( rc != 0 ) {
rc = -EINVAL;
goto fail3;
}
/* Allocate a new file for the new endpoint */
rc = onload_alloc_file(new_thr, new_s->b.bufid, priv->_filp->f_flags,
priv->fd_type, &old_ep->alien_ref);
if( rc != 0 )
goto fail4;
ci_assert(old_ep->alien_ref);
/* Copy F_SETOWN_EX, F_SETSIG to the new file */
#ifdef F_SETOWN_EX
rcu_read_lock();
__f_setown(old_ep->alien_ref->_filp, priv->_filp->f_owner.pid,
priv->_filp->f_owner.pid_type, 1);
rcu_read_unlock();
#endif
old_ep->alien_ref->_filp->f_owner.signum = priv->_filp->f_owner.signum;
old_ep->alien_ref->_filp->f_flags |= priv->_filp->f_flags & O_NONBLOCK;
/* Move os_socket from one ep to another */
if( tcp_helper_endpoint_set_aflags(new_ep, OO_THR_EP_AFLAG_ATTACHED) &
OO_THR_EP_AFLAG_ATTACHED ) {
fput(old_ep->alien_ref->_filp);
rc = -EBUSY;
goto fail2; /* state & filters are cleared by fput() */
}
/********* Point of no return **********/
ci_wmb();
priv->fd_type = CI_PRIV_TYPE_ALIEN_EP;
priv->_filp->f_op = &linux_tcp_helper_fops_alien;
ci_wmb();
oo_file_moved(priv);
/* Read all already-arrived packets after the filters move but before
* copying of the receive queue. */
ci_netif_poll(&old_thr->netif);
tcp_helper_endpoint_move_filters_post(old_ep, new_ep);
ci_assert( efab_file_move_supported(&old_thr->netif, old_s));
/* There's a gap between un-registering the old ep, and registering the
* the new. However, the notifications shouldn't be in use for sockets
* that are in a state that can be moved, so this shouldn't be a problem.
*/
if( old_ep->os_sock_pt.whead ) {
pollwait_register = 1;
efab_tcp_helper_os_pollwait_unregister(old_ep);
}
ci_assert_equal(new_ep->os_socket, NULL);
new_ep->os_socket = oo_file_ref_xchg(&old_ep->os_socket, NULL);
ci_assert_equal(old_ep->os_socket, NULL);
if( pollwait_register )
efab_tcp_helper_os_pollwait_register(new_ep);
ci_bit_clear(&new_s->b.sb_aflags, CI_SB_AFLAG_ORPHAN_BIT);
if( new_s->b.state == CI_TCP_ESTABLISHED )
CI_TCP_STATS_INC_CURR_ESTAB(alien_ni);
/* Copy recv queue */
if( new_s->b.state & CI_TCP_STATE_TCP ) {
ci_tcp_state *new_ts = SOCK_TO_TCP(new_s);
ci_tcp_state *old_ts = SOCK_TO_TCP(old_s);
int i;
/* Stop timers */
ci_ip_timer_clear(&old_thr->netif, &old_ts->kalive_tid);
ci_ip_timer_clear(&old_thr->netif, &old_ts->delack_tid);
efab_ip_queue_copy(alien_ni, &new_ts->recv1,
&old_thr->netif, &old_ts->recv1);
efab_ip_queue_copy(alien_ni, &new_ts->recv2,
&old_thr->netif, &old_ts->recv2);
new_ts->recv1_extract = new_ts->recv1.head;
/* Drop reorder buffer */
ci_ip_queue_init(&new_ts->rob);
new_ts->dsack_block = OO_PP_INVALID;
new_ts->dsack_start = new_ts->dsack_end = 0;
for( i = 0; i <= CI_TCP_SACK_MAX_BLOCKS; i++ )
new_ts->last_sack[i] = OO_PP_NULL;
}
else {
/* There should not be any recv q, but drop it to be sure */
ci_udp_recv_q_init(&SOCK_TO_UDP(new_s)->recv_q);
}
/* Old stack can be unlocked */
old_s->b.sb_flags |= CI_SB_FLAG_MOVED;
ci_netif_unlock(&old_thr->netif);
ci_assert( efab_file_move_supported(alien_ni, new_s) );
/* Move done: poll for any new data. */
ci_netif_poll(alien_ni);
if( new_s->b.state & CI_TCP_STATE_TCP ) {
ci_tcp_state *new_ts = SOCK_TO_TCP(new_s);
/* Timers setup: delack, keepalive */
if( (new_ts->acks_pending & CI_TCP_ACKS_PENDING_MASK) > 0)
ci_tcp_timeout_delack(alien_ni, new_ts);
ci_tcp_kalive_reset(alien_ni, new_ts);
}
/* Old ep: we are done. */
ci_bit_set(&old_s->b.sb_aflags, CI_SB_AFLAG_MOVED_AWAY_BIT);
old_s->b.moved_to_stack_id = alien_ni->state->stack_id;
old_s->b.moved_to_sock_id = new_s->b.bufid;
if( ! list_empty(&priv->_filp->f_ep_links) )
ci_bit_set(&old_s->b.sb_aflags, CI_SB_AFLAG_MOVED_AWAY_IN_EPOLL_BIT);
ci_sock_unlock(&old_thr->netif, &old_s->b);
ci_sock_unlock(alien_ni, &new_s->b);
ci_assert(ci_netif_is_locked(alien_ni));
OO_DEBUG_TCPH(ci_log("%s: -> [%d:%d] %s", __func__,
new_thr->id, new_s->b.bufid,
ci_tcp_state_str(new_s->b.state)));
return 0;
fail4:
/* We clear the filters from the new ep.
* For now, we do not need to re-insert old filters because hw filters
* are alredy here (in case of accepted socket) or not needed.
* We have not removed old sw filters yet. */
tcp_helper_endpoint_move_filters_undo(old_ep, new_ep);
fail3:
if( new_s->b.state & CI_TCP_STATE_TCP )
ci_tcp_state_free(alien_ni, SOCK_TO_TCP(new_s));
else
ci_udp_state_free(alien_ni, SOCK_TO_UDP(new_s));
fail2:
ci_netif_unlock(alien_ni);
fail1:
ci_netif_unlock(&old_thr->netif);
fail1_ni_unlocked:
ci_sock_unlock(&old_thr->netif, &old_s->b);
OO_DEBUG_TCPH(ci_log("%s: rc=%d", __func__, rc));
return rc;
}
